Re: [PATCH] PCI/ATS: Allow to enable ATS on VFs even if it is not enabled on PF

[Bjorn Helgaas]

On Thu, Feb 16, 2023 at 11:12:24AM +0000, Jean-Philippe Brucker wrote:
> On Wed, Feb 15, 2023 at 02:57:26PM -0600, Bjorn Helgaas wrote:
> > On Sun, Feb 12, 2023 at 08:14:48PM +0200, Leon Romanovsky wrote:
> > > On Wed, Feb 08, 2023 at 10:43:21AM -0800, Ganapatrao Kulkarni wrote:
> > > > As per PCIe specification(section 10.5), If a VF implements an
> > > > ATS capability, its associated PF must implement an ATS capability.
> > > > The ATS Capabilities in VFs and their associated PFs are permitted to
> > > > be enabled independently.
> > > > Also, it states that the Smallest Translation Unit (STU) for VFs must be
> > > > hardwired to Zero and the associated PF's value applies to VFs STU.
> > > > 
> > > > The current code allows to enable ATS on VFs only if it is already
> > > > enabled on associated PF, which is not necessary as per the specification.
> > > > 
> > > > It is only required to have valid STU programmed on PF to enable
> > > > ATS on VFs. Adding code to write the first VFs STU to a PF's STU
> > > > when PFs ATS is not enabled.
> > >
> > > Can you please add here quotes from the spec and its version? I don't see
> > > anything like this in my version of PCIe specification.
> > 
> > See PCIe r6.0, sec 10.5.1.
> > 
> > > > Signed-off-by: Ganapatrao Kulkarni <gankulkarni@xxxxxxxxxxxxxxxxxxxxxx>
> > > > ---
> > > >  drivers/pci/ats.c | 15 +++++++++++----
> > > >  1 file changed, 11 insertions(+), 4 deletions(-)
> > > > 
> > > > diff --git a/drivers/pci/ats.c b/drivers/pci/ats.c
> > > > index f9cc2e10b676..a97ec67201d1 100644
> > > > --- a/drivers/pci/ats.c
> > > > +++ b/drivers/pci/ats.c
> > > > @@ -67,13 +67,20 @@ int pci_enable_ats(struct pci_dev *dev, int ps)
> > > >  	if (ps < PCI_ATS_MIN_STU)
> > > >  		return -EINVAL;
> > > >  
> > > > -	/*
> > > > -	 * Note that enabling ATS on a VF fails unless it's already enabled
> > > > -	 * with the same STU on the PF.
> > > > -	 */
> > > >  	ctrl = PCI_ATS_CTRL_ENABLE;
> > > >  	if (dev->is_virtfn) {
> > > >  		pdev = pci_physfn(dev);
> > > > +
> > > > +		if (!pdev->ats_enabled &&
> > > > +				(pdev->ats_stu < PCI_ATS_MIN_STU)) {
> > > > +			u16 ctrl2;
> > > > +
> > > > +			/* Associated PF's STU value applies to VFs. */
> > > > +			pdev->ats_stu = ps;
> > > > +			ctrl2 = PCI_ATS_CTRL_STU(pdev->ats_stu - PCI_ATS_MIN_STU);
> > > > +			pci_write_config_word(pdev, pdev->ats_cap + PCI_ATS_CTRL, ctrl2);
> > > > +		}
> > 
> > For reference, it is this way because of edc90fee916b ("PCI: Allocate
> > ATS struct during enumeration").  The rationale was that since the PF
> > STU applies to all VFs, we should require that the PF STU be
> > programmed before enabling ATS on any of the VFs.
> > 
> > This patch relaxes that so the PF STU would be set either by (a)
> > enabling ATS on the PF or (b) enabling ATS on the first VF.
> > 
> > This looks racy because theoretically drivers for VF A and VF B could
> > independently call pci_enable_ats() with different IOMMU page sizes,
> > and we don't know which will get there first.
> > 
> > Most callers supply a compile-time constant (PAGE_SHIFT or
> > VTD_PAGE_SHIFT), so it won't matter.  arm_smmu_enable_ats() is
> > fancier, but I *assume* it would still supply the same IOMMU page size
> > for all VFs of a given PF.
> > 
> > But it's still kind of ugly to call pci_enable_ats(dev_A) and have it
> > muck with the configuration of dev_B.  Maybe we should configure the
> > PF STU (without enabling ATS) at enumeration-time in pci_ats_init()?
> > Is there some way to get the IOMMU page size at that time?
> 
> Not really, on Arm the supported page sizes are discovered when probing
> the SMMU registers, which may happen later than enumeration, during module
> loading.
>
> What this patch is trying to solve is:
> * want the PF to bypass SMMU translation, and the VF to undergo SMMU
>   translation (in order to assign it to a VM)
> * SMMU forbids enabling ATS for a configuration that bypasses translation.
>   So the PF ATS capability must be left disabled.
> 
> For this situation I wonder if we could do: the SMMU driver, seeing that
> the PF is configured to bypass, calls a new function "pci_configure_ats()"
> instead of pci_enable_ats(), which would only set the STU but leave the
> cap disabled. Then when setting up translation for the VF, the SMMU driver
> calls pci_enable_ats() as usual, which sees the PF's STU set appropriately
> and succeeds.

Seems reasonable.

It's weird to me that the SMMU is between PCI and memory, but the
driver seems to insert itself in the middle after PCI enumeration.
And maybe even after some PCI device driver binding?  But I guess if
you arrange for the SMMU driver to configure the PF before the VF
driver gets started, that's all we need from a PCI perspective.

Bjorn